Hacker News Reader

Not sure what's scarier, the faces or the message from Chtulu in the text!

Unless these examples were cherry picked rare fails, I'm surprised it's so bad when there are a few of other generative models out there including ones by OpenAI which do human faces almost perfectly. Unless they're really throttling processing power used?

Wow this is not production ready!

What is microsoft doing?

Look at midjourney 5's absolute perfection as an example.

The models at the moment don't do well on scenes where the faces are small and far away. I assume it's due to training on 256px by 256px images and not being able to capture enough detail in that, so the gens look horrible like this.

At present, when using Stable Diffusion, you can address this as part of your workflow by fixing the faces with inpainting / upscaling. This can also be automated in a scriptable way using the Unprompted extension.

This is worse than SD 1